Date of Award
5-2-2022
Degree Name
Doctor of Philosophy (PhD)
Degree Type
Dissertation
Abstract
Technological developments in external beam radiation therapy (RT) made in the past decades have been largely focused on the tight integration of imaging systems and therapy delivery systems in order to enable the precise delivery of therapeutic doses of radiation to internal, complexly shaped targets. Along this line, the adoption of x-ray computed tomography (CT) in the RT workflow and the combination of volumetric cone-beam CT with linacs in the clinical setting brought the practice of image-guided RT (IGRT) to the forefront. Practical limitations related to the poor soft tissue contrast of these x-ray imaging modalities prompted the adoption of magnetic resonance imaging (MRI) as a secondary modality in the IGRT workflow, allowing for the delineation of targets and critical structures at sites throughout the body. More recently, the development and growing use of combination MR-linac platforms has placed a focus on MR-only RT workflows in which MRI is the sole imaging modality used for treatment guidance and planning. In addition to the improved soft tissue contrast, the functional imaging capabilities and capacity for real-time image guidance of MRI have opened the door to adaptive RT (ART) applications enabled by MRI guidance in which treatments are adapted based on observed changes in anatomy or biology. These nascent MRI-guided ART workflows face several limitations, however. First, treatment planning in the adaptive setting relies on the representation of anatomy as shown in simulation scans often acquired weeks prior to treatment delivery. When the anatomy observed at treatment differs drastically from that shown at simulation, a time-intensive re-planning process must be undertaken. Second, the trade-off between spatial resolution and acquisition time means achieving high-resolution MRI scans requires scanning periods or breath-holds that may be infeasible, introducing detrimental motion artifacts. Finally, MRI-guided workflows face a technical hurdle stemming from the requirement for electron density information for dose calculations, which may be derived directly from CT images. The inclusion of CT simulation scans in the MRI-guided RT setting brings concerns related not only to extra imaging dose but also challenges with multi-modality image registration that can give rise to geometric errors that persist throughout treatment. This dissertation presents optimizations in the MRI-guided ART workflow with a focus on these challenges. First, a treatment planning strategy for pancreatic cancer cases that is robust to inter-fraction variations of primary critical structures is presented as a means of simplifying the daily adaptive planning workflow and improving target coverage in adapted fractions. Following this, a deep learning (DL)-based approach to MRI super-resolution reconstruction that enables the use of fast, low-resolution scans for guidance through fourfold upscaling is presented. Finally, the challenge of achieving electron density information through DL-based synthetic CT (sCT) reconstruction is explored in three contexts: 1) a parameter-efficient network architecture is explored in the context of sCT reconstruction in the breast for use in the low-data setting; 2) an approach to paired-data sCT reconstruction in the abdomen that handles the challenge posed by the variable presence of intestinal gas in corresponding MRI and CT scans is presented; and 3) two distinct approaches to improving the sCT outputs of an unpaired-data framework---a cascade ensemble approach and a personalized training strategy originally designed for use in the paired-data setting---are explored in the context of the sCT reconstruction task in the male pelvis. Ultimately, the methods presented in this dissertation represent improvements at three vital stages of the MRI-guided RT workflow that could potentially enable an effective and practical approach to MR-only ART.
Language
English (en)
Chair
Hong Chen